static LLVMValueRef get_wave_id_in_tg(struct si_shader_context *ctx)
{
- return si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
+ return si_unpack_param(ctx, ctx->merged_wave_info, 24, 4);
}
static LLVMValueRef get_tgsize(struct si_shader_context *ctx)
{
- return si_unpack_param(ctx, ctx->param_merged_wave_info, 28, 4);
+ return si_unpack_param(ctx, ctx->merged_wave_info, 28, 4);
}
static LLVMValueRef get_thread_id_in_tg(struct si_shader_context *ctx)
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef tmp;
tmp = LLVMBuildMul(builder, get_wave_id_in_tg(ctx),
- LLVMConstInt(ctx->ac.i32, 64, false), "");
+ LLVMConstInt(ctx->ac.i32, ctx->ac.wave_size, false), "");
return LLVMBuildAdd(builder, tmp, ac_get_thread_id(&ctx->ac), "");
}
static LLVMValueRef ngg_get_vtx_cnt(struct si_shader_context *ctx)
{
- return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
- LLVMConstInt(ctx->ac.i32, 12, false),
- LLVMConstInt(ctx->ac.i32, 9, false),
- false);
+ return si_unpack_param(ctx, ctx->gs_tg_info, 12, 9);
}
static LLVMValueRef ngg_get_prim_cnt(struct si_shader_context *ctx)
{
- return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
- LLVMConstInt(ctx->ac.i32, 22, false),
- LLVMConstInt(ctx->ac.i32, 9, false),
- false);
+ return si_unpack_param(ctx, ctx->gs_tg_info, 22, 9);
}
-static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
+static LLVMValueRef ngg_get_ordered_id(struct si_shader_context *ctx)
{
- LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
- ctx->param_rw_buffers);
-
- return ac_build_load_to_sgpr(&ctx->ac, buf_ptr,
- LLVMConstInt(ctx->i32, GFX10_GS_QUERY_BUF, false));
+ return si_unpack_param(ctx, ctx->gs_tg_info, 0, 12);
}
-/* Send GS Alloc Req message from the first wave of the group to SPI.
- * Message payload is:
- * - bits 0..10: vertices in group
- * - bits 12..22: primitives in group
- */
-static void build_sendmsg_gs_alloc_req(struct si_shader_context *ctx,
- LLVMValueRef vtx_cnt,
- LLVMValueRef prim_cnt)
+static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
{
- LLVMBuilderRef builder = ctx->ac.builder;
- LLVMValueRef tmp;
-
- tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
- ac_build_ifcc(&ctx->ac, tmp, 5020);
+ LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
- tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->ac.i32, 12, false),"");
- tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
- ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_ALLOC_REQ, tmp);
-
- ac_build_endif(&ctx->ac, 5020);
+ return ac_build_load_to_sgpr(&ctx->ac, buf_ptr,
+ LLVMConstInt(ctx->i32, GFX10_GS_QUERY_BUF, false));
}
struct ngg_prim {
LLVMValueRef isnull;
LLVMValueRef index[3];
LLVMValueRef edgeflag[3];
+ LLVMValueRef passthrough;
};
static void build_export_prim(struct si_shader_context *ctx,
struct ac_export_args args;
LLVMValueRef tmp;
- tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
- args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");
-
- for (unsigned i = 0; i < prim->num_vertices; ++i) {
- tmp = LLVMBuildShl(builder, prim->index[i],
- LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
- args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
- tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
- tmp = LLVMBuildShl(builder, tmp,
- LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
- args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
+ if (prim->passthrough) {
+ args.out[0] = prim->passthrough;
+ } else {
+ tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
+ args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");
+
+ for (unsigned i = 0; i < prim->num_vertices; ++i) {
+ tmp = LLVMBuildShl(builder, prim->index[i],
+ LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
+ args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
+ tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
+ tmp = LLVMBuildShl(builder, tmp,
+ LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
+ args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
+ }
}
args.out[0] = LLVMBuildBitCast(builder, args.out[0], ctx->ac.f32, "");
ac_build_export(&ctx->ac, &args);
}
+static void build_streamout_vertex(struct si_shader_context *ctx,
+ LLVMValueRef *so_buffer, LLVMValueRef *wg_offset_dw,
+ unsigned stream, LLVMValueRef offset_vtx,
+ LLVMValueRef vertexptr)
+{
+ struct tgsi_shader_info *info = &ctx->shader->selector->info;
+ struct pipe_stream_output_info *so = &ctx->shader->selector->so;
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef offset[4] = {};
+ LLVMValueRef tmp;
+
+ for (unsigned buffer = 0; buffer < 4; ++buffer) {
+ if (!wg_offset_dw[buffer])
+ continue;
+
+ tmp = LLVMBuildMul(builder, offset_vtx,
+ LLVMConstInt(ctx->i32, so->stride[buffer], false), "");
+ tmp = LLVMBuildAdd(builder, wg_offset_dw[buffer], tmp, "");
+ offset[buffer] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->i32, 2, false), "");
+ }
+
+ for (unsigned i = 0; i < so->num_outputs; ++i) {
+ if (so->output[i].stream != stream)
+ continue;
+
+ unsigned reg = so->output[i].register_index;
+ struct si_shader_output_values out;
+ out.semantic_name = info->output_semantic_name[reg];
+ out.semantic_index = info->output_semantic_index[reg];
+
+ for (unsigned comp = 0; comp < 4; comp++) {
+ tmp = ac_build_gep0(&ctx->ac, vertexptr,
+ LLVMConstInt(ctx->i32, 4 * reg + comp, false));
+ out.values[comp] = LLVMBuildLoad(builder, tmp, "");
+ out.vertex_stream[comp] =
+ (info->output_streams[reg] >> (2 * comp)) & 3;
+ }
+
+ si_emit_streamout_output(ctx, so_buffer, offset, &so->output[i], &out);
+ }
+}
+
+struct ngg_streamout {
+ LLVMValueRef num_vertices;
+
+ /* per-thread data */
+ LLVMValueRef prim_enable[4]; /* i1 per stream */
+ LLVMValueRef vertices[3]; /* [N x i32] addrspace(LDS)* */
+
+ /* Output */
+ LLVMValueRef emit[4]; /* per-stream emitted primitives (only valid for used streams) */
+};
+
+/**
+ * Build streamout logic.
+ *
+ * Implies a barrier.
+ *
+ * Writes number of emitted primitives to gs_ngg_scratch[4:8].
+ *
+ * Clobbers gs_ngg_scratch[8:].
+ */
+static void build_streamout(struct si_shader_context *ctx,
+ struct ngg_streamout *nggso)
+{
+ struct tgsi_shader_info *info = &ctx->shader->selector->info;
+ struct pipe_stream_output_info *so = &ctx->shader->selector->so;
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
+ LLVMValueRef tid = get_thread_id_in_tg(ctx);
+ LLVMValueRef tmp, tmp2;
+ LLVMValueRef i32_2 = LLVMConstInt(ctx->i32, 2, false);
+ LLVMValueRef i32_4 = LLVMConstInt(ctx->i32, 4, false);
+ LLVMValueRef i32_8 = LLVMConstInt(ctx->i32, 8, false);
+ LLVMValueRef so_buffer[4] = {};
+ unsigned max_num_vertices = 1 + (nggso->vertices[1] ? 1 : 0) +
+ (nggso->vertices[2] ? 1 : 0);
+ LLVMValueRef prim_stride_dw[4] = {};
+ LLVMValueRef prim_stride_dw_vgpr = LLVMGetUndef(ctx->i32);
+ int stream_for_buffer[4] = { -1, -1, -1, -1 };
+ unsigned bufmask_for_stream[4] = {};
+ bool isgs = ctx->type == PIPE_SHADER_GEOMETRY;
+ unsigned scratch_emit_base = isgs ? 4 : 0;
+ LLVMValueRef scratch_emit_basev = isgs ? i32_4 : ctx->i32_0;
+ unsigned scratch_offset_base = isgs ? 8 : 4;
+ LLVMValueRef scratch_offset_basev = isgs ? i32_8 : i32_4;
+
+ ac_llvm_add_target_dep_function_attr(ctx->main_fn, "amdgpu-gds-size", 256);
+
+ /* Determine the mapping of streamout buffers to vertex streams. */
+ for (unsigned i = 0; i < so->num_outputs; ++i) {
+ unsigned buf = so->output[i].output_buffer;
+ unsigned stream = so->output[i].stream;
+ assert(stream_for_buffer[buf] < 0 || stream_for_buffer[buf] == stream);
+ stream_for_buffer[buf] = stream;
+ bufmask_for_stream[stream] |= 1 << buf;
+ }
+
+ for (unsigned buffer = 0; buffer < 4; ++buffer) {
+ if (stream_for_buffer[buffer] == -1)
+ continue;
+
+ assert(so->stride[buffer]);
+
+ tmp = LLVMConstInt(ctx->i32, so->stride[buffer], false);
+ prim_stride_dw[buffer] = LLVMBuildMul(builder, tmp, nggso->num_vertices, "");
+ prim_stride_dw_vgpr = ac_build_writelane(
+ &ctx->ac, prim_stride_dw_vgpr, prim_stride_dw[buffer],
+ LLVMConstInt(ctx->i32, buffer, false));
+
+ so_buffer[buffer] = ac_build_load_to_sgpr(
+ &ctx->ac, buf_ptr,
+ LLVMConstInt(ctx->i32, SI_VS_STREAMOUT_BUF0 + buffer, false));
+ }
+
+ tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->i32_0, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5200);
+ {
+ LLVMTypeRef gdsptr = LLVMPointerType(ctx->i32, AC_ADDR_SPACE_GDS);
+ LLVMValueRef gdsbase = LLVMBuildIntToPtr(builder, ctx->i32_0, gdsptr, "");
+
+ /* Advance the streamout offsets in GDS. */
+ LLVMValueRef offsets_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
+ LLVMValueRef generated_by_stream_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
+
+ tmp = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5210);
+ {
+ if (isgs) {
+ tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid);
+ tmp = LLVMBuildLoad(builder, tmp, "");
+ } else {
+ tmp = ac_build_writelane(&ctx->ac, ctx->i32_0,
+ ngg_get_prim_cnt(ctx), ctx->i32_0);
+ }
+ LLVMBuildStore(builder, tmp, generated_by_stream_vgpr);
+
+ unsigned swizzle[4];
+ int unused_stream = -1;
+ for (unsigned stream = 0; stream < 4; ++stream) {
+ if (!info->num_stream_output_components[stream]) {
+ unused_stream = stream;
+ break;
+ }
+ }
+ for (unsigned buffer = 0; buffer < 4; ++buffer) {
+ if (stream_for_buffer[buffer] >= 0) {
+ swizzle[buffer] = stream_for_buffer[buffer];
+ } else {
+ assert(unused_stream >= 0);
+ swizzle[buffer] = unused_stream;
+ }
+ }
+
+ tmp = ac_build_quad_swizzle(&ctx->ac, tmp,
+ swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
+ tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
+
+ LLVMValueRef args[] = {
+ LLVMBuildIntToPtr(builder, ngg_get_ordered_id(ctx), gdsptr, ""),
+ tmp,
+ ctx->i32_0, // ordering
+ ctx->i32_0, // scope
+ ctx->ac.i1false, // isVolatile
+ LLVMConstInt(ctx->i32, 4 << 24, false), // OA index
+ ctx->ac.i1true, // wave release
+ ctx->ac.i1true, // wave done
+ };
+ tmp = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.ordered.add",
+ ctx->i32, args, ARRAY_SIZE(args), 0);
+
+ /* Keep offsets in a VGPR for quick retrieval via readlane by
+ * the first wave for bounds checking, and also store in LDS
+ * for retrieval by all waves later. */
+ LLVMBuildStore(builder, tmp, offsets_vgpr);
+
+ tmp2 = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
+ scratch_offset_basev, "");
+ tmp2 = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp2);
+ LLVMBuildStore(builder, tmp, tmp2);
+ }
+ ac_build_endif(&ctx->ac, 5210);
+
+ /* Determine the max emit per buffer. This is done via the SALU, in part
+ * because LLVM can't generate divide-by-multiply if we try to do this
+ * via VALU with one lane per buffer.
+ */
+ LLVMValueRef max_emit[4] = {};
+ for (unsigned buffer = 0; buffer < 4; ++buffer) {
+ if (stream_for_buffer[buffer] == -1)
+ continue;
+
+ LLVMValueRef bufsize_dw =
+ LLVMBuildLShr(builder,
+ LLVMBuildExtractElement(builder, so_buffer[buffer], i32_2, ""),
+ i32_2, "");
+
+ tmp = LLVMBuildLoad(builder, offsets_vgpr, "");
+ LLVMValueRef offset_dw =
+ ac_build_readlane(&ctx->ac, tmp,
+ LLVMConstInt(ctx->i32, buffer, false));
+
+ tmp = LLVMBuildSub(builder, bufsize_dw, offset_dw, "");
+ tmp = LLVMBuildUDiv(builder, tmp, prim_stride_dw[buffer], "");
+
+ tmp2 = LLVMBuildICmp(builder, LLVMIntULT, bufsize_dw, offset_dw, "");
+ max_emit[buffer] = LLVMBuildSelect(builder, tmp2, ctx->i32_0, tmp, "");
+ }
+
+ /* Determine the number of emitted primitives per stream and fixup the
+ * GDS counter if necessary.
+ *
+ * This is complicated by the fact that a single stream can emit to
+ * multiple buffers (but luckily not vice versa).
+ */
+ LLVMValueRef emit_vgpr = ctx->i32_0;
+
+ for (unsigned stream = 0; stream < 4; ++stream) {
+ if (!info->num_stream_output_components[stream])
+ continue;
+
+ tmp = LLVMBuildLoad(builder, generated_by_stream_vgpr, "");
+ LLVMValueRef generated =
+ ac_build_readlane(&ctx->ac, tmp,
+ LLVMConstInt(ctx->i32, stream, false));
+
+ LLVMValueRef emit = generated;
+ for (unsigned buffer = 0; buffer < 4; ++buffer) {
+ if (stream_for_buffer[buffer] == stream)
+ emit = ac_build_umin(&ctx->ac, emit, max_emit[buffer]);
+ }
+
+ emit_vgpr = ac_build_writelane(&ctx->ac, emit_vgpr, emit,
+ LLVMConstInt(ctx->i32, stream, false));
+
+ /* Fixup the offset using a plain GDS atomic if we overflowed. */
+ tmp = LLVMBuildICmp(builder, LLVMIntULT, emit, generated, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5221); /* scalar branch */
+ tmp = LLVMBuildLShr(builder,
+ LLVMConstInt(ctx->i32, bufmask_for_stream[stream], false),
+ ac_get_thread_id(&ctx->ac), "");
+ tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5222);
+ {
+ tmp = LLVMBuildSub(builder, generated, emit, "");
+ tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
+ tmp2 = LLVMBuildGEP(builder, gdsbase, &tid, 1, "");
+ LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpSub, tmp2, tmp,
+ LLVMAtomicOrderingMonotonic, false);
+ }
+ ac_build_endif(&ctx->ac, 5222);
+ ac_build_endif(&ctx->ac, 5221);
+ }
+
+ tmp = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5225);
+ {
+ tmp = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
+ scratch_emit_basev, "");
+ tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp);
+ LLVMBuildStore(builder, emit_vgpr, tmp);
+ }
+ ac_build_endif(&ctx->ac, 5225);
+ }
+ ac_build_endif(&ctx->ac, 5200);
+
+ /* Determine the workgroup-relative per-thread / primitive offset into
+ * the streamout buffers */
+ struct ac_wg_scan primemit_scan[4] = {};
+
+ if (isgs) {
+ for (unsigned stream = 0; stream < 4; ++stream) {
+ if (!info->num_stream_output_components[stream])
+ continue;
+
+ primemit_scan[stream].enable_exclusive = true;
+ primemit_scan[stream].op = nir_op_iadd;
+ primemit_scan[stream].src = nggso->prim_enable[stream];
+ primemit_scan[stream].scratch =
+ ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
+ LLVMConstInt(ctx->i32, 12 + 8 * stream, false));
+ primemit_scan[stream].waveidx = get_wave_id_in_tg(ctx);
+ primemit_scan[stream].numwaves = get_tgsize(ctx);
+ primemit_scan[stream].maxwaves = 8;
+ ac_build_wg_scan_top(&ctx->ac, &primemit_scan[stream]);
+ }
+ }
+
+ ac_build_s_barrier(&ctx->ac);
+
+ /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
+ LLVMValueRef wgoffset_dw[4] = {};
+
+ {
+ LLVMValueRef scratch_vgpr;
+
+ tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ac_get_thread_id(&ctx->ac));
+ scratch_vgpr = LLVMBuildLoad(builder, tmp, "");
+
+ for (unsigned buffer = 0; buffer < 4; ++buffer) {
+ if (stream_for_buffer[buffer] >= 0) {
+ wgoffset_dw[buffer] = ac_build_readlane(
+ &ctx->ac, scratch_vgpr,
+ LLVMConstInt(ctx->i32, scratch_offset_base + buffer, false));
+ }
+ }
+
+ for (unsigned stream = 0; stream < 4; ++stream) {
+ if (info->num_stream_output_components[stream]) {
+ nggso->emit[stream] = ac_build_readlane(
+ &ctx->ac, scratch_vgpr,
+ LLVMConstInt(ctx->i32, scratch_emit_base + stream, false));
+ }
+ }
+ }
+
+ /* Write out primitive data */
+ for (unsigned stream = 0; stream < 4; ++stream) {
+ if (!info->num_stream_output_components[stream])
+ continue;
+
+ if (isgs) {
+ ac_build_wg_scan_bottom(&ctx->ac, &primemit_scan[stream]);
+ } else {
+ primemit_scan[stream].result_exclusive = tid;
+ }
+
+ tmp = LLVMBuildICmp(builder, LLVMIntULT,
+ primemit_scan[stream].result_exclusive,
+ nggso->emit[stream], "");
+ tmp = LLVMBuildAnd(builder, tmp, nggso->prim_enable[stream], "");
+ ac_build_ifcc(&ctx->ac, tmp, 5240);
+ {
+ LLVMValueRef offset_vtx =
+ LLVMBuildMul(builder, primemit_scan[stream].result_exclusive,
+ nggso->num_vertices, "");
+
+ for (unsigned i = 0; i < max_num_vertices; ++i) {
+ tmp = LLVMBuildICmp(builder, LLVMIntULT,
+ LLVMConstInt(ctx->i32, i, false),
+ nggso->num_vertices, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5241);
+ build_streamout_vertex(ctx, so_buffer, wgoffset_dw,
+ stream, offset_vtx, nggso->vertices[i]);
+ ac_build_endif(&ctx->ac, 5241);
+ offset_vtx = LLVMBuildAdd(builder, offset_vtx, ctx->i32_1, "");
+ }
+ }
+ ac_build_endif(&ctx->ac, 5240);
+ }
+}
+
+static unsigned ngg_nogs_vertex_size(struct si_shader *shader)
+{
+ unsigned lds_vertex_size = 0;
+
+ /* The edgeflag is always stored in the last element that's also
+ * used for padding to reduce LDS bank conflicts. */
+ if (shader->selector->so.num_outputs)
+ lds_vertex_size = 4 * shader->selector->info.num_outputs + 1;
+ if (shader->selector->info.writes_edgeflag)
+ lds_vertex_size = MAX2(lds_vertex_size, 1);
+
+ return lds_vertex_size;
+}
+
+/**
+ * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
+ * for the vertex outputs.
+ */
+static LLVMValueRef ngg_nogs_vertex_ptr(struct si_shader_context *ctx,
+ LLVMValueRef vtxid)
+{
+ /* The extra dword is used to avoid LDS bank conflicts. */
+ unsigned vertex_size = ngg_nogs_vertex_size(ctx->shader);
+ LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, vertex_size);
+ LLVMTypeRef pai32 = LLVMPointerType(ai32, AC_ADDR_SPACE_LDS);
+ LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, ctx->esgs_ring, pai32, "");
+ return LLVMBuildGEP(ctx->ac.builder, tmp, &vtxid, 1, "");
+}
+
/**
* Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
*/
LLVMValueRef *addrs)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
- struct tgsi_shader_info *info = &ctx->shader->selector->info;
- struct si_shader_output_values *outputs = NULL;
+ struct si_shader_selector *sel = ctx->shader->selector;
+ struct tgsi_shader_info *info = &sel->info;
+ struct si_shader_output_values outputs[PIPE_MAX_SHADER_OUTPUTS];
LLVMBuilderRef builder = ctx->ac.builder;
- struct lp_build_if_state if_state;
- LLVMValueRef tmp;
+ LLVMValueRef tmp, tmp2;
assert(!ctx->shader->is_gs_copy_shader);
assert(info->num_outputs <= max_outputs);
- outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
+ LLVMValueRef vertex_ptr = NULL;
+
+ if (sel->so.num_outputs || sel->info.writes_edgeflag)
+ vertex_ptr = ngg_nogs_vertex_ptr(ctx, get_thread_id_in_tg(ctx));
for (unsigned i = 0; i < info->num_outputs; i++) {
outputs[i].semantic_name = info->output_semantic_name[i];
outputs[i].semantic_index = info->output_semantic_index[i];
- /* This is used only by streamout. */
for (unsigned j = 0; j < 4; j++) {
- outputs[i].values[j] =
- LLVMBuildLoad(builder,
- addrs[4 * i + j],
- "");
outputs[i].vertex_stream[j] =
(info->output_streams[i] >> (2 * j)) & 3;
+
+ /* TODO: we may store more outputs than streamout needs,
+ * but streamout performance isn't that important.
+ */
+ if (sel->so.num_outputs) {
+ tmp = ac_build_gep0(&ctx->ac, vertex_ptr,
+ LLVMConstInt(ctx->i32, 4 * i + j, false));
+ tmp2 = LLVMBuildLoad(builder, addrs[4 * i + j], "");
+ tmp2 = ac_to_integer(&ctx->ac, tmp2);
+ LLVMBuildStore(builder, tmp2, tmp);
+ }
+ }
+
+ /* Store the edgeflag at the end (if streamout is enabled) */
+ if (info->output_semantic_name[i] == TGSI_SEMANTIC_EDGEFLAG &&
+ sel->info.writes_edgeflag) {
+ LLVMValueRef edgeflag = LLVMBuildLoad(builder, addrs[4 * i], "");
+ /* The output is a float, but the hw expects a 1-bit integer. */
+ edgeflag = LLVMBuildFPToUI(ctx->ac.builder, edgeflag, ctx->i32, "");
+ edgeflag = ac_build_umin(&ctx->ac, edgeflag, ctx->i32_1);
+
+ tmp = LLVMConstInt(ctx->i32, ngg_nogs_vertex_size(ctx->shader) - 1, 0);
+ tmp = ac_build_gep0(&ctx->ac, vertex_ptr, tmp);
+ LLVMBuildStore(builder, edgeflag, tmp);
}
}
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
- LLVMValueRef prims_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 8, 8);
- LLVMValueRef vtx_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 0, 8);
- LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
- ac_get_thread_id(&ctx->ac), prims_in_wave, "");
- LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
- ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
+ LLVMValueRef is_gs_thread = si_is_gs_thread(ctx);
+ LLVMValueRef is_es_thread = si_is_es_thread(ctx);
LLVMValueRef vtxindex[] = {
- si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 0, 16),
- si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 16, 16),
- si_unpack_param(ctx, ctx->param_gs_vtx23_offset, 0, 16),
+ si_unpack_param(ctx, ctx->gs_vtx01_offset, 0, 16),
+ si_unpack_param(ctx, ctx->gs_vtx01_offset, 16, 16),
+ si_unpack_param(ctx, ctx->gs_vtx23_offset, 0, 16),
};
/* Determine the number of vertices per primitive. */
LLVMValueRef num_vertices_val;
if (ctx->type == PIPE_SHADER_VERTEX) {
- if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
+ if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD]) {
/* Blits always use axis-aligned rectangles with 3 vertices. */
num_vertices = 3;
num_vertices_val = LLVMConstInt(ctx->i32, 3, 0);
} else {
/* Extract OUTPRIM field. */
- tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 2, 2);
+ tmp = si_unpack_param(ctx, ctx->vs_state_bits, 2, 2);
num_vertices_val = LLVMBuildAdd(builder, tmp, ctx->i32_1, "");
num_vertices = 3; /* TODO: optimize for points & lines */
}
num_vertices_val = LLVMConstInt(ctx->i32, num_vertices, false);
}
- /* TODO: streamout */
+ /* Streamout */
+ LLVMValueRef emitted_prims = NULL;
+
+ if (sel->so.num_outputs) {
+ struct ngg_streamout nggso = {};
+
+ nggso.num_vertices = num_vertices_val;
+ nggso.prim_enable[0] = is_gs_thread;
+
+ for (unsigned i = 0; i < num_vertices; ++i)
+ nggso.vertices[i] = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
+
+ build_streamout(ctx, &nggso);
+ emitted_prims = nggso.emit[0];
+ }
+
+ LLVMValueRef user_edgeflags[3] = {};
- /* TODO: primitive culling */
+ if (sel->info.writes_edgeflag) {
+ /* Streamout already inserted the barrier, so don't insert it again. */
+ if (!sel->so.num_outputs)
+ ac_build_s_barrier(&ctx->ac);
- build_sendmsg_gs_alloc_req(ctx, ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
+ ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
+ /* Load edge flags from ES threads and store them into VGPRs in GS threads. */
+ for (unsigned i = 0; i < num_vertices; i++) {
+ tmp = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
+ tmp2 = LLVMConstInt(ctx->i32, ngg_nogs_vertex_size(ctx->shader) - 1, 0);
+ tmp = ac_build_gep0(&ctx->ac, tmp, tmp2);
+ tmp = LLVMBuildLoad(builder, tmp, "");
+ tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+
+ user_edgeflags[i] = ac_build_alloca_undef(&ctx->ac, ctx->i1, "");
+ LLVMBuildStore(builder, tmp, user_edgeflags[i]);
+ }
+ ac_build_endif(&ctx->ac, 5400);
+ }
+
+ /* Copy Primitive IDs from GS threads to the LDS address corresponding
+ * to the ES thread of the provoking vertex.
+ */
+ if (ctx->type == PIPE_SHADER_VERTEX &&
+ ctx->shader->key.mono.u.vs_export_prim_id) {
+ /* Streamout and edge flags use LDS. Make it idle, so that we can reuse it. */
+ if (sel->so.num_outputs || sel->info.writes_edgeflag)
+ ac_build_s_barrier(&ctx->ac);
+
+ ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
+ /* Extract the PROVOKING_VTX_INDEX field. */
+ LLVMValueRef provoking_vtx_in_prim =
+ si_unpack_param(ctx, ctx->vs_state_bits, 4, 2);
+
+ /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
+ LLVMValueRef indices = ac_build_gather_values(&ctx->ac, vtxindex, 3);
+ LLVMValueRef provoking_vtx_index =
+ LLVMBuildExtractElement(builder, indices, provoking_vtx_in_prim, "");
+
+ LLVMBuildStore(builder, ac_get_arg(&ctx->ac, ctx->args.gs_prim_id),
+ ac_build_gep0(&ctx->ac, ctx->esgs_ring, provoking_vtx_index));
+ ac_build_endif(&ctx->ac, 5400);
+ }
+
+ ac_build_sendmsg_gs_alloc_req(&ctx->ac, get_wave_id_in_tg(ctx),
+ ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
/* Update query buffer */
- tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
- ac_build_ifcc(&ctx->ac, tmp, 5030);
- tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac), ctx->ac.i32_0, "");
- ac_build_ifcc(&ctx->ac, tmp, 5031);
- {
- LLVMValueRef args[] = {
- ngg_get_prim_cnt(ctx),
- ngg_get_query_buf(ctx),
- LLVMConstInt(ctx->i32, 16, false), /* offset of stream[0].generated_primitives */
- ctx->i32_0, /* soffset */
- ctx->i32_0, /* cachepolicy */
- };
+ /* TODO: this won't catch 96-bit clear_buffer via transform feedback. */
+ if (ctx->screen->use_ngg_streamout &&
+ !info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD]) {
+ tmp = si_unpack_param(ctx, ctx->vs_state_bits, 6, 1);
+ tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
+ tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5030);
+ tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac),
+ sel->so.num_outputs ? ctx->ac.i32_1 : ctx->ac.i32_0, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5031);
+ {
+ LLVMValueRef args[] = {
+ ngg_get_prim_cnt(ctx),
+ ngg_get_query_buf(ctx),
+ LLVMConstInt(ctx->i32, 16, false), /* offset of stream[0].generated_primitives */
+ ctx->i32_0, /* soffset */
+ ctx->i32_0, /* cachepolicy */
+ };
- /* TODO: should this be 64-bit atomics? */
- ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
- ctx->i32, args, 5, 0);
+ if (sel->so.num_outputs) {
+ args[0] = ac_build_writelane(&ctx->ac, args[0], emitted_prims, ctx->i32_1);
+ args[2] = ac_build_writelane(&ctx->ac, args[2],
+ LLVMConstInt(ctx->i32, 24, false), ctx->i32_1);
+ }
+
+ /* TODO: should this be 64-bit atomics? */
+ ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
+ ctx->i32, args, 5, 0);
+ }
+ ac_build_endif(&ctx->ac, 5031);
+ ac_build_endif(&ctx->ac, 5030);
+ ac_build_endif(&ctx->ac, 5029);
}
- ac_build_endif(&ctx->ac, 5031);
- ac_build_endif(&ctx->ac, 5030);
/* Export primitive data to the index buffer. Format is:
* - bits 0..8: index 0
* TODO: culling depends on the primitive type, so can have some
* interaction here.
*/
- lp_build_if(&if_state, &ctx->gallivm, is_gs_thread);
+ ac_build_ifcc(&ctx->ac, is_gs_thread, 6001);
{
struct ngg_prim prim = {};
- prim.num_vertices = num_vertices;
- prim.isnull = ctx->ac.i1false;
- memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
-
- for (unsigned i = 0; i < num_vertices; ++i) {
- tmp = LLVMBuildLShr(builder, ctx->abi.gs_invocation_id,
- LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
- prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
+ if (gfx10_is_ngg_passthrough(ctx->shader)) {
+ prim.passthrough = ac_get_arg(&ctx->ac, ctx->gs_vtx01_offset);
+ } else {
+ prim.num_vertices = num_vertices;
+ prim.isnull = ctx->ac.i1false;
+ memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
+
+ for (unsigned i = 0; i < num_vertices; ++i) {
+ if (ctx->type != PIPE_SHADER_VERTEX) {
+ prim.edgeflag[i] = ctx->i1false;
+ continue;
+ }
+
+ tmp = LLVMBuildLShr(builder,
+ ac_get_arg(&ctx->ac, ctx->args.gs_invocation_id),
+ LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
+ prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
+
+ if (sel->info.writes_edgeflag) {
+ tmp2 = LLVMBuildLoad(builder, user_edgeflags[i], "");
+ prim.edgeflag[i] = LLVMBuildAnd(builder, prim.edgeflag[i],
+ tmp2, "");
+ }
+ }
}
build_export_prim(ctx, &prim);
}
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 6001);
/* Export per-vertex data (positions and parameters). */
- lp_build_if(&if_state, &ctx->gallivm, is_es_thread);
+ ac_build_ifcc(&ctx->ac, is_es_thread, 6002);
{
unsigned i;
}
}
- /* TODO: Vertex shaders have to get PrimitiveID from GS VGPRs. */
- if (ctx->type == PIPE_SHADER_TESS_EVAL &&
- ctx->shader->key.mono.u.vs_export_prim_id) {
+ if (ctx->shader->key.mono.u.vs_export_prim_id) {
outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
outputs[i].semantic_index = 0;
- outputs[i].values[0] = ac_to_float(&ctx->ac, si_get_primitive_id(ctx, 0));
+
+ if (ctx->type == PIPE_SHADER_VERTEX) {
+ /* Wait for GS stores to finish. */
+ ac_build_s_barrier(&ctx->ac);
+
+ tmp = ac_build_gep0(&ctx->ac, ctx->esgs_ring,
+ get_thread_id_in_tg(ctx));
+ outputs[i].values[0] = LLVMBuildLoad(builder, tmp, "");
+ } else {
+ assert(ctx->type == PIPE_SHADER_TESS_EVAL);
+ outputs[i].values[0] = si_get_primitive_id(ctx, 0);
+ }
+
+ outputs[i].values[0] = ac_to_float(&ctx->ac, outputs[i].values[0]);
for (unsigned j = 1; j < 4; j++)
outputs[i].values[j] = LLVMGetUndef(ctx->f32);
si_llvm_export_vs(ctx, outputs, i);
}
- lp_build_endif(&if_state);
-
- FREE(outputs);
+ ac_build_endif(&ctx->ac, 6002);
}
static LLVMValueRef
return ngg_gs_vertex_ptr(ctx, vertexidx);
}
+static LLVMValueRef
+ngg_gs_get_emit_output_ptr(struct si_shader_context *ctx, LLVMValueRef vertexptr,
+ unsigned out_idx)
+{
+ LLVMValueRef gep_idx[3] = {
+ ctx->ac.i32_0, /* implied C-style array */
+ ctx->ac.i32_0, /* first struct entry */
+ LLVMConstInt(ctx->ac.i32, out_idx, false),
+ };
+ return LLVMBuildGEP(ctx->ac.builder, vertexptr, gep_idx, 3, "");
+}
+
+static LLVMValueRef
+ngg_gs_get_emit_primflag_ptr(struct si_shader_context *ctx, LLVMValueRef vertexptr,
+ unsigned stream)
+{
+ LLVMValueRef gep_idx[3] = {
+ ctx->ac.i32_0, /* implied C-style array */
+ ctx->ac.i32_1, /* second struct entry */
+ LLVMConstInt(ctx->ac.i32, stream, false),
+ };
+ return LLVMBuildGEP(ctx->ac.builder, vertexptr, gep_idx, 3, "");
+}
+
void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx,
unsigned stream,
LLVMValueRef *addrs)
const struct si_shader_selector *sel = ctx->shader->selector;
const struct tgsi_shader_info *info = &sel->info;
LLVMBuilderRef builder = ctx->ac.builder;
- struct lp_build_if_state if_state;
LLVMValueRef tmp;
const LLVMValueRef vertexidx =
LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
- lp_build_if(&if_state, &ctx->gallivm, can_emit);
+ ac_build_ifcc(&ctx->ac, can_emit, 9001);
const LLVMValueRef vertexptr =
ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
continue;
LLVMValueRef out_val = LLVMBuildLoad(builder, addrs[4 * i + chan], "");
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implied C-style array */
- ctx->ac.i32_0, /* first entry of struct */
- LLVMConstInt(ctx->ac.i32, out_idx, false),
- };
- LLVMValueRef ptr = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
-
out_val = ac_to_integer(&ctx->ac, out_val);
- LLVMBuildStore(builder, out_val, ptr);
+ LLVMBuildStore(builder, out_val,
+ ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx));
}
}
assert(out_idx * 4 == sel->gsvs_vertex_size);
const LLVMValueRef iscompleteprim =
LLVMBuildICmp(builder, LLVMIntUGE, curverts, tmp, "");
+ /* Since the geometry shader emits triangle strips, we need to
+ * track which primitive is odd and swap vertex indices to get
+ * the correct vertex order.
+ */
+ LLVMValueRef is_odd = ctx->i1false;
+ if (stream == 0 && u_vertices_per_prim(sel->gs_output_prim) == 3) {
+ tmp = LLVMBuildAnd(builder, curverts, ctx->i32_1, "");
+ is_odd = LLVMBuildICmp(builder, LLVMIntEQ, tmp, ctx->i32_1, "");
+ }
+
tmp = LLVMBuildAdd(builder, curverts, ctx->ac.i32_1, "");
LLVMBuildStore(builder, tmp, ctx->gs_curprim_verts[stream]);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implied C-style array */
- ctx->ac.i32_1, /* second struct entry */
- LLVMConstInt(ctx->ac.i32, stream, false),
- };
- const LLVMValueRef primflagptr =
- LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
-
+ /* The per-vertex primitive flag encoding:
+ * bit 0: whether this vertex finishes a primitive
+ * bit 1: whether the primitive is odd (if we are emitting triangle strips)
+ */
tmp = LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i8, "");
- LLVMBuildStore(builder, tmp, primflagptr);
+ tmp = LLVMBuildOr(builder, tmp,
+ LLVMBuildShl(builder,
+ LLVMBuildZExt(builder, is_odd, ctx->ac.i8, ""),
+ ctx->ac.i8_1, ""), "");
+ LLVMBuildStore(builder, tmp, ngg_gs_get_emit_primflag_ptr(ctx, vertexptr, stream));
tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 9001);
}
void gfx10_ngg_gs_emit_prologue(struct si_shader_context *ctx)
LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
tmp = ngg_gs_emit_vertex_ptr(ctx, gsthread, vertexidx);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implied C-style array */
- ctx->ac.i32_1, /* second entry of struct */
- LLVMConstInt(ctx->ac.i32, stream, false),
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- LLVMBuildStore(builder, i8_0, tmp);
+ LLVMBuildStore(builder, i8_0, ngg_gs_get_emit_primflag_ptr(ctx, tmp, stream));
ac_build_endloop(&ctx->ac, 5100);
}
LLVMValueRef numprims =
LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
- numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, 64);
+ numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, ctx->ac.wave_size);
tmp = LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(&ctx->ac), ctx->i32_0, "");
ac_build_ifcc(&ctx->ac, tmp, 5105);
ac_build_endif(&ctx->ac, 5105);
}
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
ac_build_s_barrier(&ctx->ac);
const LLVMValueRef tid = get_thread_id_in_tg(ctx);
LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);
- /* TODO: streamout */
+ /* Streamout */
+ if (sel->so.num_outputs) {
+ struct ngg_streamout nggso = {};
- tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->i32, 4, false), "");
- ac_build_ifcc(&ctx->ac, tmp, 5110);
- {
- LLVMValueRef offset;
- tmp = tid;
- if (sel->so.num_outputs)
- tmp = LLVMBuildAnd(builder, tmp, LLVMConstInt(ctx->i32, 3, false), "");
- offset = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 32, false), "");
- if (sel->so.num_outputs) {
- tmp = LLVMBuildLShr(builder, tid, LLVMConstInt(ctx->i32, 2, false), "");
- tmp = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 8, false), "");
- offset = LLVMBuildAdd(builder, offset, tmp, "");
+ nggso.num_vertices = LLVMConstInt(ctx->i32, verts_per_prim, false);
+
+ LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tid);
+ for (unsigned stream = 0; stream < 4; ++stream) {
+ if (!info->num_stream_output_components[stream])
+ continue;
+
+ tmp = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, vertexptr, stream), "");
+ tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+ tmp2 = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
+ nggso.prim_enable[stream] = LLVMBuildAnd(builder, tmp, tmp2, "");
}
- tmp = LLVMBuildLoad(builder, ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid), "");
- LLVMValueRef args[] = {
- tmp,
- ngg_get_query_buf(ctx),
- offset,
- LLVMConstInt(ctx->i32, 16, false), /* soffset */
- ctx->i32_0, /* cachepolicy */
- };
- ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
- ctx->i32, args, 5, 0);
+ for (unsigned i = 0; i < verts_per_prim; ++i) {
+ tmp = LLVMBuildSub(builder, tid,
+ LLVMConstInt(ctx->i32, verts_per_prim - i - 1, false), "");
+ tmp = ngg_gs_vertex_ptr(ctx, tmp);
+ nggso.vertices[i] = ac_build_gep0(&ctx->ac, tmp, ctx->i32_0);
+ }
+
+ build_streamout(ctx, &nggso);
+ }
+
+ /* Write shader query data. */
+ if (ctx->screen->use_ngg_streamout) {
+ tmp = si_unpack_param(ctx, ctx->vs_state_bits, 6, 1);
+ tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
+ unsigned num_query_comps = sel->so.num_outputs ? 8 : 4;
+ tmp = LLVMBuildICmp(builder, LLVMIntULT, tid,
+ LLVMConstInt(ctx->i32, num_query_comps, false), "");
+ ac_build_ifcc(&ctx->ac, tmp, 5110);
+ {
+ LLVMValueRef offset;
+ tmp = tid;
+ if (sel->so.num_outputs)
+ tmp = LLVMBuildAnd(builder, tmp, LLVMConstInt(ctx->i32, 3, false), "");
+ offset = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 32, false), "");
+ if (sel->so.num_outputs) {
+ tmp = LLVMBuildLShr(builder, tid, LLVMConstInt(ctx->i32, 2, false), "");
+ tmp = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 8, false), "");
+ offset = LLVMBuildAdd(builder, offset, tmp, "");
+ }
+
+ tmp = LLVMBuildLoad(builder, ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid), "");
+ LLVMValueRef args[] = {
+ tmp,
+ ngg_get_query_buf(ctx),
+ offset,
+ LLVMConstInt(ctx->i32, 16, false), /* soffset */
+ ctx->i32_0, /* cachepolicy */
+ };
+ ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
+ ctx->i32, args, 5, 0);
+ }
+ ac_build_endif(&ctx->ac, 5110);
+ ac_build_endif(&ctx->ac, 5109);
}
- ac_build_endif(&ctx->ac, 5110);
/* TODO: culling */
/* Determine vertex liveness. */
- LLVMValueRef vertliveptr = lp_build_alloca(&ctx->gallivm, ctx->ac.i1, "vertexlive");
+ LLVMValueRef vertliveptr = ac_build_alloca(&ctx->ac, ctx->ac.i1, "vertexlive");
tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
ac_build_ifcc(&ctx->ac, tmp, 5120);
/* Load primitive liveness */
tmp = ngg_gs_vertex_ptr(ctx, primidx);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_0, /* stream 0 */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- tmp = LLVMBuildLoad(builder, tmp, "");
+ tmp = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 0), "");
const LLVMValueRef primlive =
LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
* there are 4 or more contiguous null primitives in the export
* (in the common case of single-dword prim exports).
*/
- build_sendmsg_gs_alloc_req(ctx, vertlive_scan.result_reduce, num_emit_threads);
+ ac_build_sendmsg_gs_alloc_req(&ctx->ac, get_wave_id_in_tg(ctx),
+ vertlive_scan.result_reduce, num_emit_threads);
/* Setup the reverse vertex compaction permutation. We re-use stream 1
* of the primitive liveness flags, relying on the fact that each
ac_build_ifcc(&ctx->ac, vertlive, 5130);
{
tmp = ngg_gs_vertex_ptr(ctx, vertlive_scan.result_exclusive);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_1, /* stream 1 */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
tmp2 = LLVMBuildTrunc(builder, tid, ctx->ac.i8, "");
- LLVMBuildStore(builder, tmp2, tmp);
+ LLVMBuildStore(builder, tmp2, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 1));
}
ac_build_endif(&ctx->ac, 5130);
tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
ac_build_ifcc(&ctx->ac, tmp, 5140);
{
+ LLVMValueRef flags;
struct ngg_prim prim = {};
prim.num_vertices = verts_per_prim;
tmp = ngg_gs_vertex_ptr(ctx, tid);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_0, /* primflag */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- tmp = LLVMBuildLoad(builder, tmp, "");
- prim.isnull = LLVMBuildICmp(builder, LLVMIntEQ, tmp,
- LLVMConstInt(ctx->ac.i8, 0, false), "");
+ flags = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 0), "");
+ prim.isnull = LLVMBuildNot(builder, LLVMBuildTrunc(builder, flags, ctx->i1, ""), "");
for (unsigned i = 0; i < verts_per_prim; ++i) {
prim.index[i] = LLVMBuildSub(builder, vertlive_scan.result_exclusive,
prim.edgeflag[i] = ctx->ac.i1false;
}
+ /* Geometry shaders output triangle strips, but NGG expects triangles.
+ * We need to change the vertex order for odd triangles to get correct
+ * front/back facing by swapping 2 vertex indices, but we also have to
+ * keep the provoking vertex in the same place.
+ *
+ * If the first vertex is provoking, swap index 1 and 2.
+ * If the last vertex is provoking, swap index 0 and 1.
+ */
+ if (verts_per_prim == 3) {
+ LLVMValueRef is_odd = LLVMBuildLShr(builder, flags, ctx->ac.i8_1, "");
+ is_odd = LLVMBuildTrunc(builder, is_odd, ctx->i1, "");
+ LLVMValueRef flatshade_first =
+ LLVMBuildICmp(builder, LLVMIntEQ,
+ si_unpack_param(ctx, ctx->vs_state_bits, 4, 2),
+ ctx->i32_0, "");
+
+ struct ngg_prim in = prim;
+ prim.index[0] = LLVMBuildSelect(builder, flatshade_first,
+ in.index[0],
+ LLVMBuildSelect(builder, is_odd,
+ in.index[1], in.index[0], ""), "");
+ prim.index[1] = LLVMBuildSelect(builder, flatshade_first,
+ LLVMBuildSelect(builder, is_odd,
+ in.index[2], in.index[1], ""),
+ LLVMBuildSelect(builder, is_odd,
+ in.index[0], in.index[1], ""), "");
+ prim.index[2] = LLVMBuildSelect(builder, flatshade_first,
+ LLVMBuildSelect(builder, is_odd,
+ in.index[1], in.index[2], ""),
+ in.index[2], "");
+ }
+
build_export_prim(ctx, &prim);
}
ac_build_endif(&ctx->ac, 5140);
tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, vertlive_scan.result_reduce, "");
ac_build_ifcc(&ctx->ac, tmp, 5145);
{
- struct si_shader_output_values *outputs = NULL;
- outputs = MALLOC(info->num_outputs * sizeof(outputs[0]));
+ struct si_shader_output_values outputs[PIPE_MAX_SHADER_OUTPUTS];
tmp = ngg_gs_vertex_ptr(ctx, tid);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_1, /* stream 1: source data index */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- tmp = LLVMBuildLoad(builder, tmp, "");
+ tmp = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 1), "");
tmp = LLVMBuildZExt(builder, tmp, ctx->ac.i32, "");
const LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tmp);
unsigned out_idx = 0;
- gep_idx[1] = ctx->ac.i32_0;
for (unsigned i = 0; i < info->num_outputs; i++) {
outputs[i].semantic_name = info->output_semantic_name[i];
outputs[i].semantic_index = info->output_semantic_index[i];
for (unsigned j = 0; j < 4; j++, out_idx++) {
- gep_idx[2] = LLVMConstInt(ctx->ac.i32, out_idx, false);
- tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
+ tmp = ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx);
tmp = LLVMBuildLoad(builder, tmp, "");
outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
outputs[i].vertex_stream[j] =
}
si_llvm_export_vs(ctx, outputs, info->num_outputs);
-
- FREE(outputs);
}
ac_build_endif(&ctx->ac, 5145);
}
shader->previous_stage_sel ? shader->previous_stage_sel : gs_sel;
const enum pipe_shader_type gs_type = gs_sel->type;
const unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
- /* TODO: Specialize for known primitive type without GS. */
- const unsigned input_prim = gs_type == PIPE_SHADER_GEOMETRY ?
- gs_sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM] :
- PIPE_PRIM_TRIANGLES;
+ const unsigned input_prim = si_get_input_prim(gs_sel);
const bool use_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
const unsigned max_verts_per_prim = u_vertices_per_prim(input_prim);
/* We can't allow using the whole LDS, because GS waves compete with
* other shader stages for LDS space.
*
- * Streamout can increase the ESGS buffer size later on, so be more
- * conservative with streamout and use 4K dwords. This may be suboptimal.
- *
- * Otherwise, use the limit of 7K dwords. The reason is that we need
- * to leave some headroom for the max_esverts increase at the end.
- *
* TODO: We should really take the shader's internal LDS use into
* account. The linker will fail if the size is greater than
* 8K dwords.
*/
- const unsigned max_lds_size = (gs_sel->so.num_outputs ? 4 : 7) * 1024 - 128;
+ const unsigned max_lds_size = 8 * 1024 - 768;
const unsigned target_lds_size = max_lds_size;
unsigned esvert_lds_size = 0;
unsigned gsprim_lds_size = 0;
/* All these are per subgroup: */
bool max_vert_out_per_gs_instance = false;
- unsigned max_esverts_base = 256;
+ unsigned max_esverts_base = 128;
unsigned max_gsprims_base = 128; /* default prim group size clamp */
/* Hardware has the following non-natural restrictions on the value
esvert_lds_size = es_sel->esgs_itemsize / 4;
gsprim_lds_size = (gs_sel->gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
} else {
- /* TODO: This needs to be adjusted once LDS use for compaction
- * after culling is implemented. */
+ /* VS and TES. */
+ /* LDS size for passing data from ES to GS. */
+ esvert_lds_size = ngg_nogs_vertex_size(shader);
+
+ /* LDS size for passing data from GS to ES.
+ * GS stores Primitive IDs into LDS at the address corresponding
+ * to the ES thread of the provoking vertex. All ES threads
+ * load and export PrimitiveID for their thread.
+ */
+ if (gs_sel->type == PIPE_SHADER_VERTEX &&
+ shader->key.mono.u.vs_export_prim_id)
+ esvert_lds_size = MAX2(esvert_lds_size, 1);
}
unsigned max_gsprims = max_gsprims_base;
/* Round up towards full wave sizes for better ALU utilization. */
if (!max_vert_out_per_gs_instance) {
- const unsigned wavesize = 64;
+ const unsigned wavesize = gs_sel->screen->ge_wave_size;
unsigned orig_max_esverts;
unsigned orig_max_gsprims;
do {